Framework Update | AG2 v1.0.2

🌏 中文版

Version Info

Why This Release Matters

The focus of this release isn't piling on new features — it's upgrading "multiple agent systems calling each other" from "it works" to "it works securely." The previous release (v1.0.1) had just pinned the mcp dependency to <2 in response to MCP v2.0 breaking changes. This release follows up by closing two security gaps in A2A (Agent-to-Agent) communication: agent cards move from plaintext to signature verification, and the transport layer gains gRPC TLS. At the same time, AG2 agents can now bi-directionally expose themselves as ACP (Agent Client Protocol) services, letting external clients (e.g. IDEs or other agent frameworks) drive them via HTTP/WebSocket — not just the other way around. Cross-referencing the watchlist item tracking AG2's "A2A support," this is the next step on that line.

Notable Changes

• Bi-directional ACP support: AG2 agents can be exposed as ACP agents, letting remote clients drive them over HTTP or WebSocket; ACP protocol upgraded to 0.12 with Kilo Code support and model selection added -> Previously AG2 mostly "actively" called external tools/agents; now it can also be passively invoked as a service by other systems
• A2A agent card signing and verification: Outgoing agent cards are signed; incoming agent cards are verified for origin -> Prevents malicious or forged agent cards from being trusted
• A2A gRPC TLS transport: Agent-to-agent communication gains a gRPC TLS transport layer -> Cross-network agent calls can be encrypted without wrapping your own proxy
• A2A user-defined extensions: Users can define custom A2A extensions -> Attach custom fields/behavior on top of the standard protocol
• LiveAgent adds ElevenLabs voice provider: Another provider option for real-time voice interaction
• First community extensions: Tenki (isolated execution sandbox), TealTiger (deterministic governance middleware)
• Developer tooling: Agent composition exposed as a read-only public property, middleware gains describe() for introspection, tool types consolidated into ag2.tools.types

Breaking Changes

None in this release.

Migration Guide

Upgrade directly — no code changes required:

pip install --upgrade ag2==1.0.2

To start using new features, here's an example:

# Expose an existing AG2 agent as an ACP service for remote clients
from ag2.acp import ACPServer

server = ACPServer(agent=my_agent, transport="websocket")
server.serve()

Cross-Framework Observations

A2A interop security is something multiple frameworks are actively backfilling. LangGraph and CrewAI currently rely on external gateways to handle cross-framework agent authentication, while AG2 now bakes signature verification directly into the framework layer. Combined with the bi-directional ACP support added in the same release, AG2 simultaneously acts as an "A2A participant" and an "ACP-driven endpoint" — further ahead on the protocol interop front than most peers.

Takeaway

When following A2A (Agent-to-Agent) updates, it's easy to focus only on "can they talk to each other." This release highlights that "agent card signing and verification" is the real gate to production readiness — an agent card is essentially an identity credential, and without signature verification, anyone can forge an agent to impersonate their way into your multi-agent system.

References

• AG2 v1.0.2 — GitHub Release
• ag2ai/ag2 — GitHub